The Application of Reinforcement Learning Algorithms in UAV Flight Control Training Courses

Nowadays, traditional methods about UAV flight training field face many challenges such as high risks, high costs, and insufficient generalization. In this paper, it discusses on the application of imitation learning algorithms in UAV control training field, especially exploring three algorithms such as Behavior Cloning (BC), Inverse Reinforcement Learning (IRL), and Generative Adversarial Imitation Learning (GAIL). Research has show that: 1) Although BC is very applicable for rapid training of standardized trajectories through supervised learning, it still exists distribution shift issues; 2) IRL can carry out multi-objective optimization by obtaining reward functions, but this way brings high computational complexity; 3) GAIL could enhance adaptabilities in the complex environments by utilizing generative adversarial networks to approximate expert trajectory distributions, but it probably meets the mode collapse issues. The paper also discusses algorithm optimization strategies about different aspects such as hierarchical reward design, data augmentation, and trajectory truncation techniques. This work can provides selection advice in different scenarios, especially from the entry-level training to engineering deployment. About the future work, the paper also mentions in multi-agent collaboration, demonstration-free learning, and physical constraint embedding aspects, and it is helpful to promote the in-depth application of imitation learning in UAV autonomous control.